Autonomous multi-angle capture

Pixel and perspective geometry for the simplest multi-angle rig, a 360 camera on a pole reframed in software. An 8K (7,680 px) equirect cropped into a 1080×1920 vertical reel is native only up to about 24 mm full-frame-equivalent; 50 mm is upscaled 2.1× and 85 mm 3.5×, and a native 85 mm crop would need a ~27,000 px equirect. Perspective is set by distance: from a stick at 0.8 m a feature 10 cm nearer the lens images 14% larger, against 7.6% from the 1.4 m an 85 mm head-and-shoulders shot is taken at, with the background 1.7× larger. So portrait and compressed-background shots need a camera standing back, whatever the resolution. Upper bounds from pixel counts (real cameras resolve less); framing distances and depths are stated assumptions.

Exact search over 17,280 kits of cameras, lenses, mounts and battery packs (11,520 legal), scoring distinct focal-band × viewpoint shots (18 at most) and recording hours against carried weight. With a 4 h recording floor, exact search reaches 15 shots at 4 kg and all 18 at 5 kg, while a greedy best-gain-per-gram packer stalls at 12 from 3 kg because a lens is worthless without its body. Without a floor, variety-first search returns kits that record only ~2.2 h. Every device figure is assumed (round class-typical numbers, not datasheets); only the 100 Wh limit is sourced.

A 2-D geometric model of N cameras on a ring or a 90° front arc 3 m from a subject, scored on angle coverage and triangulation error. Cameras buy coverage, angle buys depth accuracy: a 30° gap between neighbours needs 13 cameras (15° needs 25), while centre depth error barely moves with count (0.52 mm for 3 on a ring, 0.23 mm for 16) but falls 11× as a pair’s convergence angle goes from 5° to 90° (7.3 mm to 0.64 mm). Four cameras on a front arc are within 11% of four on a ring. Idealised lower bounds with assumed noise (0.5 px) and no occlusion; novel-view quality is not modelled.

A ray-traced toy scene, two ring cameras, a depth-based warp to the midpoint, scored against the ray-traced truth. With exact depth, synthesis beats cutting to the nearest camera only up to a gap of about 30° (24.8 against 23.6 dB); at 45° it is worse (19.3 against 23.1 dB) because 8% of the target is a hole neither camera saw, though pixels it does fill stay near 29–30 dB. A bystander pulls the crossover in to about 15°, and depth error on the sources costs about 1.4 dB per percent at first (27.5 dB exact, 22.0 dB at 10%). Toy scene with smooth textures and oracle depth: read the ordering, not the decibels.

capture-novelview assumed independent depth noise of 2–5%. Here depth is measured instead: block-matching stereo on a rendered rectified pair, fed into the same two-view warp and scored against ray-traced truth at a 15° gap. Mean depth error falls from 14.9% at a 3 cm baseline to 2.7% at 50 cm, but coverage peaks at 91% near 20 cm (stereo cannot match occluded or flat pixels), and the synthesised view never beats cutting to the nearest camera: best 22.3 dB at 20 cm against 24.8 dB, with exact depth at 27.3 dB. Independent noise with the same mean error is optimistic by 1optimistic by 1–5 dBndash;7 dB, because real errors are correlated and missing depth leaves holes. Sensor noise of 3% already costs 3 dB. Toy scene, smooth textures, a classic matcher with no learned prior or hole filler.

capture-stereodepth found that synthesis from block-matching stereo depth never beat cutting to the nearest camera (22.3 against 24.8 dB at a 15° gap). Here the same holey, correlated depth is repaired two ways: filling holes with the farther of the nearest valid row neighbours (27.4 dB), and a gate that needs no ground truth, where the two warped sources disagree or a hole remains, falls back to the raw nearest camera (29.2 dB, 4.4 dB above cutting; the same on three held-out noise seeds). Gains hold from 8° to 30° gaps and at 6–50 cm baselines, shrink to nothing at 6% sensor noise, and with a bystander the gate loses to plain hole filling (26.1 against 26.3 dB). A per-pixel oracle at 32.0 dB bounds what any confidence rule could reach. Stylised toy scene; read the ordering, not the decibels.

A 2-D crowd simulation of a tripod, a wheeled robot, a wheel-legged robot and a drone holding a camera 2.5 m behind a walking creator among pedestrians who do not react to it. A drone that ignores pedestrians frames the creator 99.9% of the time at any density, but steering around people drops it to 69% at 0.1 people/m² and 39% at 0.3; a ground robot on a plaza with curbs and one ramp is stopped 42–47% of the walk, and wheel-legs hold the frame 70% of the time in a thin crowd but 4.5% in a dense one. Avoidance and framing trade off: a flat-floor wheeled robot that ignores people frames 98% of the time while passing within 0.5 m of a pedestrian about 500 times an hour. All platform speeds, powers and masses are assumed; pedestrians do not avoid the platforms, and the drone is not limited by altitude, wind or regulation.

A synthetic model of four cameras whose true angle quality changes every 2–8 s, seen through a noisy detector. At noise 0.3 a fixed 3 s rotation shows mean true quality 0.546; an offline Viterbi edit that maximises detected quality minus a cost per cut reaches 0.802 against 0.815 for the same search on true scores, while picking the best score every frame gives 0.723 with 345 cuts a minute and every shot under 1.5 s. The cut cost matters more than detector accuracy (best near 1.0, about 14 cuts a minute). Clock sync from a shared loudness envelope fails first on short clips: a 5 s clip at noise 0.25 is off by over 40 ms in 25% of trials, 15 s in 5%. No real footage, reframing or film-look grade; quality is a made-up random process.

An assumed size-to-recall model of blurring strangers' faces while filming in public. A face detector that is 50% reliable at 20 px leaves a face-on face of at least 32 px unblurred 5% of the time per frame (8% at 5 m on an 8K 360 camera); blurring the head region of any detected body instead drops it under 0.1%, because a body is 11× taller than a face. A telephoto frame resolves about 10 identifiable people at once against 2.5 for an 8K 360 view. Per-frame blur is worse than it looks: with independent misses a person filmed for 5 s at the edge of range shows unblurred in some frame 100% of the time, 25% if misses last the whole shot, and an offline edit that blurs a whole track after one detection is left with 0–5%. Size alone does not separate the tall platforms. All detector curves, sizes and angles are assumptions; no real detector or footage; blurring is not legal compliance.

A synthetic 16 h day with assumed device figures. With 6 interesting 30 s moments an hour and a trigger that misses 10% and raises 2 false alarms an hour, a standby camera with a 10 s pre-roll buffer uses 19.4 Wh against 80 Wh continuous (76% less) and captures 90% of the interesting seconds; a deep-sleep camera uses 7.7 Wh but captures 69%, because it wakes too slowly to catch the start. Coverage is set by the trigger's miss rate and the wake latency, not the battery; energy decides only for a rig (four cameras: 320 Wh continuous, 78 Wh triggered). Stylised: every device figure is assumed and the trigger is a probability, not a detector.

A synthetic 16 h day with assumed device figures and four standby cameras. With 6 thirty-second moments an hour and a 10% miss per detector, a trigger per camera uses 80 Wh (of 320 Wh continuous) but puts all four angles on only 68% of the interesting seconds, since independent misses differ by camera (≈0.9⁴). One sentinel waking the rest costs the same 80 Wh and gives 90.5%; waking all on any camera's trigger gives 100% for 92 Wh but its false-alarm energy grows with the number of detectors (281 against 126 Wh at 40 false alarms an hour). Sleeping the other cameras halves the rig to 41.5 Wh at the cost of the first 7 s of each moment. If the misses are shared, extra detectors buy nothing. Stylised: every figure assumed, not measured.

A model with assumed detector figures (a miss rate of 3% when the subject faces the camera, 40% when facing away) and a subject whose facing at each moment is von Mises distributed around a favourite direction. In a four-camera ring with concentration κ = 2, the sentinel at the favourite direction misses 8.6% of moments, the one behind the subject 34.4%, and a rotating sentinel 21.5%; the closed form matches a 40,000-moment simulation to within 0.4 points. Misses across a spread ring are negatively correlated (a subject turned from one camera faces another), so assuming independent misses, as capture-wake did, overstates the chance that every detector misses by about 2× on a ring of four; for a tight cluster of cameras the correlation is positive and the independent figure understates it about 12×. A second part replaces the assumed cosine curve with the miss measured on 54 faces in capture-detectors (YuNet at 0.6: 3% near-frontal, 5% turned; YuNet at 0.9: 46% and 84%; frontal Haar cascade: 34% and 84%): placement then matters only if the back of a head goes undetected, which is untested (best 6–7% against worst 81% for YuNet at 0.6 if so, 3–4% against 5% if not), and for a weak detector the four-camera any-of miss is 0.08–0.45 rather than 0.0007. The cosine curve and the 54-face sample (a detector’s own detections, a landmark proxy for turn) are the limits. Stylised: the cosine figures are assumed; the step figures are measured on 54 faces.

Follow-up to capture-residgate section D. On ten real LibreOffice gallery clips (0 dB white noise, 0.8 s offset, 432 windows per clip and cell), the share of delay estimates over 2 ms off on the four failing clips falls from 97% at a 1 s window to 38% at 2 s and 16% at 3 s (gamma 0.25), because two cameras with a 0.8 s offset share only W−0.8 s of audio. Gamma moves 2 s results between 38% and 76% (PHAT worst). A tone clip (space2) is not helped by more audio (52–70% aliased at 2 and 3 s); only a delay prior tighter than its 12 ms period works (oracle ±10 ms: 0% wrong; ±50 ms: 65% still aliased). The early-start failures of untie and soft (59–77% wrong when the window starts in the clip's first half second, 0% later) are partly a short shared segment; the shared audio is only 2–5 dB quieter in wrong windows, so loudness is a weak explanation. Limits: ten clips, white noise at one SNR, 3 s windows barely vary their start (one draw of content per clip), the prior is an oracle ceiling.

Synthetic speech-like audio (shared source, independent noise, optional reverb), per-window GCC-PHAT delay (100–4000 Hz) and a Theil–Sen line over 12 windows in 1 h at 40 ppm clock error. With 2 s windows the fit is below 0.02 ms from −10 dB SNR up and fails completely at −15 dB; 5 s windows hold to −15 dB. A sync made once at the start drifts one 30 fps frame (33.3 ms) in 833 s at 40 ppm (144 ms after an hour; analytic ppm × time reproduced exactly). Reverb at −10 dB direct-to-reverberant adds 4adds 4–8 msndash;5 ms (median). Not from real recordings: the audio and the exact delay are synthetic and the ppm values are swept, none sourced (no datasheet reachable).

Measured recall against pixel size for real detectors on COCO photos, replacing the assumed curve in capture-bystander. 22 confident faces and 254 labelled people from 128 photos were shrunk by scales from 1 to 0.065 and each detector asked to find them again. YuNet at its usual 0.6 score found every face of 16 px or wider and half at 10 px (fitted 50% point 9.2 px, range 8.2–10.4), not the assumed 20; at a 0.9 threshold it is 26.6 px. On an 8K 360 camera the per-frame leak among faces of at least 32 px falls from the assumed 4.8% to 0.01% (36% to 0.9% at 16 px), and rises to 7.2% at the 0.9 threshold. A YOLOX-s person detector reaches 52% at 20–32 px of height and 94% at 96–128 px (50% point 23 px, not 40); the head sits in the top quarter to third of its box. Section F: by how far the face is turned from the camera, YuNet at 0.9 falls from 54% to 16% recall (54 faces) and the frontal Haar cascade from 66% to 16%, while YuNet at 0.6 holds 95%+; the turn is a landmark proxy and back-of-head is not covered. A pilot: the reference faces are easy by construction (frontal, confident), the photos are shrunk not distant, and n is small, so these are optimistic bounds; the identifiable size is still assumed.

capture-follow found that a repulsion-field controller trades framing for distance: at 0.1 people/m² a drone that ignores pedestrians frames 99.8% of the time with 523 contacts (within 0.5 m) per hour, and 67.8% with 161 once it avoids them. Here a sampler that predicts pedestrians 3 s ahead and picks a velocity that goes round and rejoins is run on the same crowds (8 seeds). Framing is not recovered at equal settings: the planner at a 1.5 m radius frames the drone 58.0% of the time (repulsion 67.8%) but with 36 contacts an hour, not 161. What it changes is the trade-off: on the drone’s framing-against-contacts curve the planner sits to the left of repulsion (89% framed at 281 contacts against 79% at 322; 58% at 36 against 58.5% at 92), and on ground platforms at 0.1 people/m² it cuts contacts from 326 to 21 (wheeled, flat plaza) with the same 40% framing. At 0.3 people/m² both controllers lose the frame (3 to 39%). Pedestrians ignore the platform and their velocities are known exactly, so the planner is an upper bound on perception. Platform figures are capture-follow’s assumed round numbers, not datasheets.

capture-followplan's planner cut a drone's close passes (within 0.5 m) at 0.1 people/m² from 523 to 36 per hour, but it was handed exact pedestrian positions and velocities. Here a constant-velocity Kalman tracker over detections with Gaussian noise (0.1 to 1.0 m) and latency (0 to 1 s) feeds it instead, on the same 8 crowds. Framing is unchanged (57 to 61%, against 58.0% with exact data); contacts are not: 45 at 0.1 m noise and no delay, 120 at 0.3 m and 0.25 s, 199 to 257 at 0.6 m, 373 at 1.0 m and 1 s. The planner beats exact-data repulsion (161) only while noise is about 0.3 m or less and latency under about 0.5 s, and that comparison favours repulsion, which was not rerun with noisy inputs. Assumed, not measured: perfect detection-to-track association, independent Gaussian noise, the filter's process noise, and capture-follow’s round-number platform figures; the noise levels are a sweep, not a detector’s datasheet.

Chooses the capture kit (11,520 kits of capture-realkit, sourced devices) together with the lightest nightly-offload storage plan from capture-storage (microSD cards, two SSD copies of a 7-day trip, parallel readers, copy within 1 h) under one weight budget. Counting storage costs 3 of 18 shots at a 4 kg budget (18 become 15), and the lightest kit reaching all 18 shots grows from 3,900 g to 4,125 g (+225 g). Scaling bitrates from 0.25× to 1.5× moves that total between 4,025 and 4,300 g, and reader speed (90, 104, 312 MB/s) changes no shot count at any budget tried. Sweeping filming hours (2–8 h/day) and trip length (3–28 days) moves the lightest 18-shot total from 3,492 g to 5,307 g; within 4 kg only the 2 h/day kit keeps all 18 shots. Sourced: the action-camera bitrate, microSD mass, bus ceilings and the device catalogue. SSD and reader masses were checked against retailer listings (32–58 g per SSD, 28–85 g reader): the lightest 18-shot kit then weighs 4,040–4,262 g and 4 kg still reaches 15 shots. Not sourced: the 360-camera, phone and mirrorless bitrates (swept), and the laptop or hub the offload needs (not counted).

A one-hidden-layer network that predicts, per pixel, how much of the raw nearest camera to blend into a stereo-depth synthesis scores 26.78 dB on 40 held-out scene configurations (new ground textures) against 26.21 dB for the best soft disagreement gate and 25.81 dB for the hard gate, ahead of the soft gate in all 40 (+0.57 dB, standard error 0.05). That recovers 0.36 of the 2.7 dB between the hard gate and the best possible per-pixel blend (soft: 0.15; ridge regression: 0.29). Outside the training ranges (gaps of 36–46°, three bystanders, heavy noise) the gain over the soft gate falls to +0.24 dB, and the matching cost adds almost nothing as a feature. Toy analytic scene, block matcher, PSNR only; every scene parameter is drawn from stated ranges, which are assumed, not sourced.

Reruns capture-audiosync's GCC-PHAT delay estimator on 10 real recordings and sound effects (installed LibreOffice gallery sounds, not copied into the repository) with simulated delay and synthetic white or wind-like noise. With 2 s windows the share of windows more than one 30 fps frame off is 0.00 on synthetic audio at −10 dB SNR but 0.53 on average over the real clips (0.00 for applause, 0.95 for the worst), and 0.26 vs 0.00 at 0 dB. Partial whitening (dividing the cross-spectrum by |C|^γ) cuts the mean failure share from 0.52 to 0.28 at −10 dB with γ = 0.25. Limits: 10 sound-effect clips, not field audio; simulated delay and noise; no reverb, where PHAT is expected to win; trials within a clip overlap.

Reruns capture-syncedit with every GCC window computed on 10 real recorded clips (the LibreOffice sound gallery, via capture-realaudio) instead of synthetic audio. 4 cameras, a 120 s Viterbi edit at a random point of a 1 h shoot, clocks uniform in ±40 ppm, a Theil–Sen fit over 12 windows of 2 s, 12 seeds per clip. The synthetic result (no damaged cuts down to −10 dB) does not carry over: the mean share of cuts jumping over one 30 fps frame is 0.34 at 0 dB and 0.64 at −10 dB, and four clips lose 62–97% of cuts at 0 dB, while applause, roll and romans stay at 0 down to −10 dB. Partial whitening (gamma 0.25 instead of PHAT) cuts the 0 dB mean to 0.014 but only to 0.48 at −10 dB. Limits: effects rather than field audio, simulated delay, noise and clocks (±40 ppm is swept, not sourced), event rate and the one-frame criterion assumed, gamma chosen in-sample on the same clips, no viewer test.

Re-runs the capture-kit search (11,520 kits, 6,912 legal) on a catalogue of real devices whose weights, battery energies and runtimes come from manufacturer and retailer pages read 2026-10-21 (Insta360 X4, DJI Osmo Action 4, iPhone 15 Pro Max, Sony a7C II with three lenses, Peak Design tripod, two Anker banks). With a 4 h recording floor, exact search reaches all 18 band × viewpoint shots at 4 kg (the assumed catalogue needed 5 kg) and a greedy packer stalls at 12, as before: the lens-needs-a-body complementarity survives real figures. Sweeping the figures no source gave (phone recording draw, selfie-pole, mini-tripod and mast masses, within ±30%) shows the ranking is robust (greedy is below exact in 77–99% of draws from 1.5 kg up) but the exact variety at a given budget is not: at 4 kg it falls from 18 to as low as 8, so the mount masses, which no source gave, matter more than any camera. Not sourced: those mass and draw figures, the iPhone recording draw (Apple publishes playback only), cell voltages used to turn mAh into Wh, and which bands each device covers.

Follow-up to capture-syncgate. Drops GCC windows more than 2–20 ms off a Theil-Sen clock line (up to 3 rounds), a threshold in time rather than in peak sharpness, over 10 real LibreOffice gallery clips (4 cameras, 12 windows, ±40 ppm, 12 seeds). At 0 dB SNR with gamma 0.25 the share of cuts jumping over a frame falls from 1.4% to 0.9% for any tolerance in the range, mostly from one clip (untie). At −10 dB it helps little (46.1% vs 47.6% ungated) where the PSR gate gets 37.8%; combined 37.4%. It does not fix the held-out gamma-choice failure from capture-syncgate (9.9%). Why trimming drops so many windows: for space2 66% of windows are period aliases (a 12.1 ms near-tone; errors at multiples of the period, 0% under a control period), wallewal 12% aliases and 14% gross, untie and soft 25–34% gross early-start failures of unknown cause. Limits: ten clips from one gallery, synthetic delay and white noise, assumed clock rates, small grids, assumed one-frame criterion.

Adds simulated reverberation (exponentially decaying Gaussian tails, RT60 0–1.0 s, direct-to-reverberant ratio +5 to −5 dB, Schroeder's statistical model) to capture-realaudio's 10 real clips and tests GCC delay estimators for camera audio sync. Across 20 cells (10 rooms, SNR −5 and +5 dB, 200 windows each), PHAT is the worst fixed weighting in 19 of 20 cells: at RT60 0.6 s, DRR 0 dB, SNR +5 dB it puts 0.26 of 2 s windows more than one 30 fps frame off, against 0.09 for partial whitening with γ = 0.5 and 0.20 for plain correlation. Choosing γ per window by peak-to-sidelobe ratio beats PHAT but not the best fixed γ; an oracle selector would roughly halve the failures again. Limits: statistical rooms rather than measured impulse responses, white noise, sound-effect clips, simulated delay.

A ramped weight on how much the nearest raw camera replaces a stereo-depth synthesis, driven by disagreement between the two warps, beats the hard switch in every condition tried by 0.1–0.5 dB (29.5 against 29.3 dB at a 15° gap on held-out seeds) and leaves about a quarter fewer false edges, but closes only about a tenth of the 2.6 dB gap to a per-pixel oracle. The stereo matching cost is a worse confidence signal than warp disagreement: gating on it is level with plain hole filling and worse under a bystander or sensor noise. Stylised toy scene, parameters tuned on development seeds.

Data volume, card and SSD weight, nightly offload time and unbacked footage for a rig of 1 to 8 cameras filming 4 h a day for a week, at assumed 60 to 200 Mbit/s bitrates. Four cameras make about 860 GB a day (6 TB a week), yet their cards weigh about 8 g and a camera's battery is 440 to 980 times heavier per recording hour than its card. The binding limit is the evening: one reader copying one day to one SSD takes 2.7 h, so two SSD copies inside 1 h need 6 readers and about 450 g, half an hour is infeasible, and halving the bitrate halves the weight. Figures in the first version are assumed, not datasheets. Update 2026-10-21: action-camera bitrate (120 Mbit/s), card mass (0.25 g microSD) and card bus ceilings (UHS-I 104, UHS-II 312 MB/s) are now sourced, and SHA-256 hashing was measured at about 1,460 MB/s per core; the conclusion holds (copy time binds, not card weight; a UHS-II reader cuts the plan from 426 g to 351 g; verify while copying). The 360-camera, phone and mirrorless bitrates and SSD figures remain assumed.

Schedule study for audio sync of two free-running cameras on synthetic speech-like audio (capture-audiosync's GCC-PHAT estimator). With a 40 ms tolerance (EBU R37 lead limit, a stand-in for a cut threshold) an offset-only sync lasts 1000 s at 40 ppm, so even perfect estimates every 1800 s leave 44% of a 6 h shoot out of tolerance. A drift gate that refuses estimates a real clock (≤100 ppm) could not produce cuts that share from 35% to 22% at −12 dB SNR and from 11% to 3% at −10 dB when re-syncing every 300 s, and removes all wrongly applied syncs at −12 dB; at 900 s or longer it does nothing, since one skipped window already exceeds the tolerance. A wrong first sync is not fixable by a gate. Not real recordings: audio is synthetic, drift is a straight line at an assumed 40 ppm, and the tolerance was set for lip-sync, not for cuts between cameras.

Edit quality cannot show it (a camera is shifted in score and picture together), so this counts continuity at each cut: the jump in world time between the two cameras, and loud events inside it. 4 cameras, a 120 s Viterbi edit from capture-edit at a random point of a 1 h shoot, camera clocks uniform in ±40 ppm, sync errors from the real capture-audiosync GCC-PHAT estimator on synthetic audio (12 seeds). A sync made once and never redone leaves 50% of cuts jumping over one 30 fps frame (mean jump 40 ms); a Theil–Sen fit over 12 windows leaves none from 0 dB down to −10 dB SNR with 2 s windows (−15 dB with 5 s). Below that it flips to 69–98% of cuts jumping 0.3–0.5 s with 11–24% of cuts repeating or dropping a loud event. A threshold, not a slope. Synthetic audio, assumed event rate 0.5/s and one-frame visibility criterion; ppm range swept, not sourced; no real footage and no viewer test.

Follow-up to capture-syncslope on synthetic speech-like audio (capture-audiosync's GCC-PHAT, residual searched). Replaces its assumed 10–30 ppm rate wobble with a clock rate computed from the AT-cut crystal temperature curve in the Corning/Tele Quarz data book (a₃ = 1.05×10⁻⁴ ppm/K³) over assumed camera temperature profiles (a 25 K warm-up, a drifting day) and assumed ±5 arcmin cut-angle errors. That curve moves the relative rate only 15 to 20 ppm over 6 h, so the plain offset hold is already within a 40 ms stand-in tolerance 98–100% of the time at 300 to 1800 s re-syncs; fitting a line matters only at 3600 s (37–43% bad to 9–16%). Theil-Sen and Kalman holds tie with least squares there. They differ only on the assumed 30 ppm wobble: a Kalman filter tuned to fast rate wander (20 ppm per √h) keeps 900 s re-syncs to 0.9% bad where every other fit gives 27% and the offset hold 13%, while at 1800 s every hold fails. Base ppm, temperature profiles, cut-angle errors and wobble are assumed; no real recording.

Coarse-to-fine audio sync on synthetic speech-like audio: one 10 s GCC-PHAT window gives the start offset (0.8 s), then every later window on camera B is cut that much earlier so short windows overlap fully. Over 1 h at 40 ppm, 0.5 s windows hold within one 30 fps frame from −12 dB SNR where plain 0.5 s windows never work; 1 s windows go from failing up to −5 dB to working at −15 dB; with reverb at 0 dB SNR 1 s windows go from 0/20 to 20/20 seeds. At −20 dB all fail. Not real recordings: audio, delay and the 40 ppm are synthetic or assumed; the plain baseline loses much of its overlap by construction.

Follow-up to capture-realedit. Adds a peak-to-sidelobe (PSR) gate that drops GCC windows whose peak is not distinct before the Theil-Sen fit, and picks gamma and the gate threshold leave-one-clip-out over 10 real LibreOffice gallery clips (4 cameras, 12 windows, ±40 ppm, 12 seeds per clip). At 0 dB SNR, gamma 0.25 with a chosen threshold leaves 1.9% of cuts jumping over a frame on held-out clips (PHAT: 34.2%); at −10 dB 39.1% (PHAT 63.7%, ungated gamma 0.25 47.6%). Choosing gamma on other clips alone fails (9.9%) because the one clip where gamma 0.5 collapses cannot be seen from the other nine. Limits: ten clips from one gallery, synthetic delay and white noise, PSR thresholds from a probe on the same clips, and the one-frame criterion is assumed.

Follow-up to capture-syncbudget on synthetic speech-like audio (capture-audiosync's GCC-PHAT, window cut with the current correction so only the residual is searched). With a steady 40 ppm clock error, holding a fitted line instead of the last offset cuts the share of a 6 h shoot beyond 40 ms (EBU R37 lead limit, a stand-in for a cut threshold) from 44% to 3.7% at one re-sync per 1800 s and from 72% to 12% at 3600 s; the remainder is the first interval. With an assumed 10 to 30 ppm rate wobble a strict line gate locks out true estimates (69–88% bad at 300 and 900 s); resetting after two rejections fixes 300 s but at 900 s and 30 ppm the plain offset hold wins (13% vs 27%). Also corrects capture-syncbudget, whose raw-window correlation let the delay outgrow the 2 s window. Wobble amplitudes are swept, not measured; no real recording.

Follow-up to capture-aliaswin, whose ±10 ms prior was centred on the true delay and gave 0% wrong windows. Here the prior is a Theil-Sen line through the camera's own earlier estimates (ten real LibreOffice clips, 2 s windows, 0 dB white noise, 0.8 s offset, ±40 ppm drift, 252 scored windows per clip). It does not reproduce the oracle: with one prior-free bootstrap window it is as bad as no prior on the four failing clips (37.4% vs 37.5% wrong) and 37% of those cameras are wrong in every scored window of the hour, because an early alias is tracked rather than corrected; five bootstrap windows reach 22.5% wrong (14% of cameras lost). Capping the half-width under the clip's period breaks clips that worked when the line has no slope yet (33.8% wrong on six clips that had 0%). Limits: constant simulated drift, one SNR, an untuned fit and half-width rule, ten clips. Added 2026-10-29: gating (re-acquire when the unrestricted GCC peak is 1.5× the in-prior peak) cuts the fraction of lost cameras on the failing clips from 36.8% to 7.6–9.7% and the wrong share to 22.1% with one bootstrap window, leaving the six good clips at 0%; the tone clip space2 is not fixed (60–81% wrong) because its aliases have equal peak heights.

Follow-up to capture-learngate. Two ways to close the gap to the best per-pixel blend were tried on the same toy scenes: four new features meant to say which warped source is wrong (signed depth order, and a cross-check that warps each source into the other's camera), and a second network that predicts a per-pixel mix of the two sources before the raw-camera blend. On 40 held-out scenes (new ground textures) the new features move the learned blend from 26.78 to 26.81 dB (+0.03, ahead in 35 of 40); the source weight moves it to 27.28 dB (+0.50 over learngate, ahead in 40 of 40), recovering 0.54 of the 2.71 dB hard-gate-to-best-blend headroom against 0.36 before. The gain comes from the second output, not the new features: with the old 12 features the source weight gets 27.24 dB. Outside the training ranges the gain shrinks to +0.04 dB over learngate. Toy analytic scene, block matcher, PSNR only; every scene parameter is drawn from stated ranges that are assumed, not sourced.